pi05-BimanualYAM-freshbase-umi100-ee20-history-t1-t2

Fresh-base Pi0.5 checkpoint trained for 12,000 optimizer steps on canonical dual-UMI data only.

• UMI dataset: brandonyang/dual-lidar-umi-independent@a95b079b2b3dc73a912ecd12967f22f825d04fa8
• task: pick up oranges and place them in the bowl
• history frames: t-1, t-2
• history state: per arm, current normalized gripper plus current-frame SE(3) log coordinates for each past pose
• state size before Pi0.5 padding: 26 dimensions
• action: H24 EE20, per-arm T_t^-1 T_(t+k), xyz + R6D rows + absolute future normalized gripper
• rotation contraction: none
• teleop data: none
• base: lerobot/pi05_base@b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba

This is a research checkpoint. The 26-D history state must be constructed with the exact training convention before inference. Hardware use still requires the normal EE-to-IK path, limits, safety checks, and operator supervision. Pin the immutable Hub revision rather than main.